← 返回
AI智能 中文

Coverify

CoVerify by MO§ES™ — the falsification instrument for the Commitment Conservation Law. Extract kernels, score Jaccard, detect ghost tokens, run model swap te...
MO§ES™ CoVerify——承诺守恒定律证伪工具。提取内核,计算Jaccard相似度,检测幽灵令牌,执行模型替换测试...
sunrisesillneversee
AI智能 clawhub v0.4.3 2 版本 100000 Key: 无需
★ 0
Stars
📥 544
下载
💾 10
安装
2
版本
#commitment#conservation#falsifiability#jaccard#latest#moses#provenance#signal#verification

概述

CoVerify by MO§ES™ — Commitment Conservation Verifier

The Claim

The Commitment Conservation Law: C(T(S)) = C(S)

Semantic commitment — the irreducible meaning encoded in a signal — is conserved

under transformation when enforcement is active. It leaks when enforcement is absent.

This is a falsifiable empirical claim. Not a framework description. Not a metaphor.

clawhub install coverify installs the falsification instrument. If the law fails

under your test conditions, the ghost token report names exactly what leaked and why.


Falsification

The law is falsified if commitment leaks under active enforcement. CoVerify is

how you test it:

# Does enforcement preserve this commitment?
python3 commitment_verify.py ghost \
  "the agent must complete the task and shall never skip verification" \
  "the agent should complete the task and can skip verification if needed"

Output: ghost token report. mustshould and shall nevercan are

HIGH-cascade leakage events — enforcement anchors softened. cascade_risk: HIGH.

The ghost_pattern fingerprint identifies the structural identity of the leak.

If the same fingerprint appears when two independent agents process the same signal,

it is not extraction variance — it is a structural flaw in the harness.

That is the falsification condition.


What It Does

Extract: Pull the hard commitment kernel C(S) from a text signal. These are

the tokens that survive compression — must, shall, never, always, require,

guarantee, and the sentences that carry them.

Compare: Jaccard similarity on two kernels. Score ≥ 0.8 = commitment conserved.

Score < 0.8 = leak or model extraction variance. The input_hash tells you which —

same hash, low Jaccard = variance. Different hashes = expected divergence.

Ghost: Step-function leakage accounting. Quantifies not just that commitment

leaked, but what leaked (the ghost_pattern fingerprint), the cascade risk

(HIGH if modal/enforcement anchors lost), and whether the leak pattern is

structural across agents.

Model Swap: Automated cross-model test. Same hashed signal through two

extraction passes. Classifies result as CONSISTENT (agreement), VARIANCE

(model subjectivity — expected), or STRUCTURAL (same ghost pattern — harness hole).


Ghost Tokens and Cascade Risk

Ghost tokens are the commitment tokens present in the original signal but

absent after transformation. The leakage model is step-function, not smooth:

cascade_risk = HIGH  if any modal/enforcement anchor leaked
cascade_risk = MEDIUM  if peripheral tokens leaked, anchors intact
cascade_risk = NONE  if no leakage

One HIGH-cascade event propagates through all downstream reasoning — the

obligation it encoded continues to be inherited by the reasoning chain,

but without the force that made it obligatory. The downstream system

looks locally healthy. The commitment is gone.

See: references/ghost-token-spec.md


Install

# Standalone verifier — the falsification instrument
clawhub install coverify

# Full constitutional governance stack (coverify is the measurement primitive)
clawhub install moses-governance

Commands

CommandWhat it does
------
python3 commitment_verify.py extract ""Extract commitment kernel + input hash
python3 commitment_verify.py compare "" ""Jaccard score + CONSERVED/VARIANCE/DIVERGED verdict
python3 commitment_verify.py ghost "" ""Step-function leakage report + ghost_pattern fingerprint
python3 commitment_verify.py verify Look up entries in audit ledger by input hash
python3 model_swap_test.py ""Cross-model structural vs. variance classification

Example: Detecting a Commitment Leak

python3 commitment_verify.py ghost \
  "Agents must always verify lineage. The system shall never skip the gate." \
  "Agents should probably verify lineage when possible."
{
  "leaked_cascade_tokens": ["must always", "shall never"],
  "cascade_risk": "HIGH",
  "cascade_note": "Modal/enforcement anchor lost. All downstream reasoning inherits softening.",
  "ghost_pattern": "a3f7c2...",
  "ghost_pattern_note": "Same ghost_pattern across two agents = structural flaw, not extraction variance."
}

Verdicts

VerdictMeaning
------
CONSERVEDJaccard ≥ 0.8 — commitment kernel survived transformation
VARIANCESame input hash, Jaccard < 0.8 — model extraction differs, not a leak
DIVERGEDDifferent inputs, Jaccard < 0.8 — commitment leaked or inputs genuinely different

What Ships

VersionWhat ships
------
v0.1extract, compare, verify — Conservation Law operational. ✓ Live.
v0.2ghost — Step-function leakage model, cascade risk, ghost_pattern fingerprint. ✓ Live.
v0.3model_swap_test — Cross-model CONSISTENT/VARIANCE/STRUCTURAL classification. ✓ Live.
v0.4Archival chain (archival.py) — pre-drop provenance. Isnad + handshake. Three-layer lineage. ⏳ Planned.

About

CoVerify is a standalone instrument from the MO§ES™ family. It implements the

Commitment Conservation Law from *"A Conservation Law for Commitment in Language

Under Transformative Compression and Recursive Application"* (Zenodo, 2026).

Every agent that installs it runs the same extraction logic tracing to the same

origin anchor. The install is a proof-of-use receipt.

See also: references/falsifiability.md, references/ghost-token-spec.md

contact@burnmydays.com · mos2es.io · GitHub

版本历史

共 2 个版本

  • v0.4.3 当前
    2026-03-29 20:39 安全 安全
  • v0.1.0
    2026-03-14 04:38

安全检测

腾讯云安全 (Keen)

安全,无风险
查看报告

腾讯云安全 (Sanbu)

安全,无风险
查看报告

🔗 相关推荐

ai-intelligence

Self-Improving + Proactive Agent

ivangdavila
自我反思+自我批评+自我学习+自组织记忆。智能体评估自身工作、发现错误并持续改进。
★ 1,362 📥 318,686
ai-intelligence

ontology

oswalpalash
类型化知识图谱,用于结构化智能体记忆与可组合技能。支持创建/查询实体(人员、项目、任务、事件、文档)及关联...
★ 714 📥 243,982
security-compliance

Moses Governance

sunrisesillneversee
MO§ES™治理框架 — AI代理的宪法执行层。支持模式、姿态、角色、SHA‑256审计链、血统托管、签名门控及提交功能。
★ 0 📥 928